Convolutional Neural Networks with Template-Based Data Augmentation for Functional Lung Image Quantification.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Pulmonology

Background:

• Accurate lung MRI segmentation and ventilation quantification are crucial for diagnosing respiratory diseases.
• Previous methods were computationally intensive, limiting their clinical application.
• Deep learning offers potential for faster and more efficient image analysis.

Purpose of the Study:

• To develop an automated deep learning pipeline for proton lung MRI segmentation.
• To enable accurate, ventilation-based quantification of lung function.
• To improve computational efficiency compared to existing methodologies.

Main Methods:

• Developed deep convolutional neural network (CNN) models using a custom multilabel Dice metric loss function.
• Implemented a novel template-based data augmentation strategy to address large data requirements.
• Trained models on 205 proton MR images and 73 functional lung MRI datasets.

Main Results:

• Achieved high accuracy for lung MRI segmentation (Dice overlap: left 0.93, right 0.94, whole 0.94).
• Demonstrated significantly reduced processing time (<1 second per subject) compared to previous methods (~30 minutes).
• Obtained accurate ventilation defect quantification (average Dice overlap 0.94), comparable to expert readers.

Conclusions:

• The proposed deep learning framework provides accurate, automated lung MRI segmentation and ventilation quantification in near real-time.
• CNNs drastically reduce processing time, showing significant potential for quantitative analysis of functional lung MRI.
• The ANTsRNet repository supports this work, offering a growing collection of deep learning architectures.